Global data center electricity consumption projected to quadruple by 2035, matching India's current usage.

What happened

Recent industry projections indicate that global data centers will consume four times more electricity by 2035 than they do today. To put this scale into perspective, new data center infrastructure built between now and 2033 is expected to draw as much power as the entire country of India currently uses. This unprecedented surge in energy demand is primarily driven by the explosive growth of generative AI and the massive, continuous compute requirements necessary for training and serving large-scale foundation models.

Technical details

Modern AI workloads are fundamentally altering server rack power densities. While traditional enterprise racks typically consumed 5 to 10 kW of power, high-density AI clusters utilizing next-generation GPUs are pushing rack densities well past 100 kW. This thermal reality requires phasing out traditional air cooling in favor of direct-to-chip liquid cooling and multi-phase immersion systems. Furthermore, the aggregate grid demand means new data centers require dedicated high-voltage substations, transmission line upgrades, and direct integration with large-scale renewable generation, grid-scale battery storage, or even small modular reactors (SMRs).

Why it matters

For engineers and system architects, power is now the primary constraint on scaling AI. The historical focus on maximizing raw FLOPs is transitioning to a strict optimization of performance-per-watt. If energy infrastructure cannot keep pace with compute deployment, the industry will face severe provisioning delays, skyrocketing operational expenditures (OPEX), and regulatory pushback due to grid instability and carbon emissions. This forces a paradigm shift where algorithmic efficiency, aggressive quantization, and specialized low-power ASICs become just as critical as raw compute capabilities.

What to watch next

Watch for hyperscalers to aggressively vertically integrate their power supply, including direct investments in nuclear, geothermal, and dedicated solar-plus-storage facilities. On the hardware side, monitor the adoption rates of silicon photonics for lower-power data transfer, as well as the development of ultra-efficient edge-compute architectures designed to offload inference from centralized, power-hungry hubs. Finally, expect regulatory bodies to introduce stricter Power Usage Effectiveness (PUE) requirements and carbon footprint caps on new facility construction.